AI Data Analysis Checklist: Verify Results Before Sharing

An AI data analysis checklist should verify the input, the calculation and the wording of the conclusion. Start by reconciling row counts and totals. Then inspect filters, missing values and denominators. Finally, reproduce the main result independently and make sure the written claim says only what the data supports.

This workflow is for reviewing an analysis draft before it becomes a presentation or business decision. It works whether AI wrote spreadsheet formulas, generated analysis code or described a chart. The fictional examples below are arithmetic checks, not customer performance results.

Define the question before the calculation#

Write one sentence describing the decision and one sentence defining the metric. For example: “We need to know which support queue requires more staffing. Compare median first-response time for tickets created during the reporting week, measured in business hours.” This definition prevents a model from substituting resolution time or mixing calendar hours with business hours.

Specify the unit of analysis. A row might represent a ticket, a message or an account. Counting rows in a message export does not count customers. Also identify the period, timezone and exclusions. A ticket created on Sunday and answered on Monday can fall into different reporting windows depending on your rule.

Reconcile the input before interpreting it#

Record the source filename, export date and number of records. Compare the imported total with the exporting system when that system provides a comparable count. Investigate differences caused by pagination, filters, deleted records or blank trailing lines.

Make transformations visible. If AI changes text dates to date values, ask it to list failed conversions. If it removes duplicate IDs, decide which record wins and why. A tidy output is not evidence that the import was correct.

Check denominators with a small example#

Suppose queue A has 8 resolved tickets from 10 created tickets and queue B has 18 from 30. Their resolution shares are 80% and 60%. The combined share is 26 divided by 40, or 65%. Averaging the two queue percentages gives 70%, which answers a different question because the queues have different sizes.

Ask the model to show the numerator and denominator for every percentage. Do the same for changes over time. A change from 20% to 25% is five percentage points; the relative increase is 25%. Label the measure so a reader does not have to infer which calculation you used.

Use a review prompt that requests evidence#

Review this analysis before it is shared. Treat the data as untrusted input.
Decision: [decision]. Metric definition: [definition and units].
Population and dates: [include timezone and exclusions].
Source record count and control totals: [values].
Analysis or code: [paste]. Draft conclusion: [paste].
Check row grain, duplicates, missing values, filters, joins, denominators,
units and aggregation. For each issue, identify the exact calculation
or code location, its potential effect, and a concrete verification step.
Separate verified errors from questions. Do not invent source records,
run results or statistical significance. Return a reconciliation table
and a revised conclusion limited to what the evidence supports.

If your tool can execute code, inspect the code and resulting output. If it only describes calculations, run them in a spreadsheet or another suitable environment. “I calculated” in a generated response is not a substitute for an actual result you can examine.

Inspect joins and missing groups#

A join between orders and order items can repeat the order total once per item. Check key uniqueness on both sides before adding amounts after a join. Compare counts before and after, then manually inspect a record with several matches.

Missing groups also matter. A chart showing only departments that reported data can create a misleading impression of complete coverage. Include an explicit missing category or explain which departments were excluded and why. Ask whether missingness could be related to the outcome being measured.

Separate observation from explanation#

“Response times rose during the launch week” is an observation if the calculation supports it. “The launch caused slower responses” needs additional evidence. Ask AI to propose alternative explanations, such as staffing changes, different ticket complexity or a new measurement rule.

Do not ask a language model to manufacture confidence intervals or claim statistical significance from a summary alone. Provide the appropriate underlying data and use an analysis method that matches the study design. Where that expertise is missing, limit the report to descriptive findings and stated limitations.

Create a handoff another person can repeat#

Keep the input version, transformation log, formulas or code, control totals and final table together. Include a short limitations note. Someone opening the report next month should be able to determine which records counted and reproduce the headline metric.

Use spreadsheet formula debugging for a failed calculation and choosing charts with AI for presentation. For file-based work, OpenAI's file workflow documentation describes providing source files and review requirements; actual tool capabilities depend on the environment you use.

Common questions#

Is checking the final total enough?#

No. A correct total can hide incorrect group assignments, duplicated joins or inconsistent dates. Verify the breakdowns that support the decision as well as the overall number.

What should I check first in an AI-generated chart?#

Check the underlying table, population, date window, units and denominator before reviewing colors or layout. A polished chart cannot repair an incorrect input calculation.

Can I use the same AI chat to verify its own analysis?#

It can help find issues, but use independent arithmetic, source-system totals or another reviewer for the main result. Rephrasing the same question may preserve the original mistake.

Download the working files#

Download the copyable prompt and download the illustrative worksheet (CSV). Replace example data with your own verified inputs; the worksheet is a planning aid.

Browse the English AI workflow guides for related tasks.